Musk is building his own chip factory

The chip wars just went vertical. Elon Musk announced Terafab, a $25 billion joint venture between Tesla, SpaceX, and xAI to manufacture custom AI chips at the Giga Texas North Campus in Austin—targeting 2nm process nodes, the most advanced in commercial production. Small-batch AI5 chips are slated for late 2026 with volume production in 2027, and 80% of the compute is earmarked for orbital AI satellites running on SpaceX’s network. Musk isn’t just designing chips anymore; he’s building the foundry, the satellites, and the cars that run on them. But while Terafab is still a blueprint, Amazon’s Trainium2 is already shipping at scale—Anthropic runs Claude on over 1 million Trainium2 chips, Project Rainier clusters 500,000 of them in one of the world’s largest AI compute installations, and OpenAI just secured 2 gigawatts of Trainium capacity as part of a $50 billion deal. Apple is testing Trainium for its own AI workloads. One company is pouring concrete on a $25 billion bet; the other already has every major lab signed to silicon that exists today. The timeline gap between announcement and production is where most chip ambitions go to die.

Custom silicon gets the headlines, but the Pentagon is locking in AI at the software layer. a memo from Deputy Secretary of Defense Steve Feinberg directed senior Pentagon leaders to accelerate Palantir’s Maven AI into the military’s permanent infrastructure—expanding it from an intelligence program into a core command-and-control system across all branches. Palantir’s Army contract is valued at up to $10 billion, the company’s market cap sits near $360 billion after its stock doubled in the past year, and Maven oversight transfers from the NGA to the Chief Digital AI Office within 30 days. The complication: Maven relies on Claude AI, and the Pentagon blacklisted Anthropic earlier this year for refusing to remove safety guardrails from weapons systems. A separate study circulating this week showed AI can now de-anonymize HackerNews users from their posts alone with 67% recall at 90% precision—a capability that barely existed a year ago. The military is formalizing AI it can’t fully control the supply chain for, while the models themselves are developing capabilities nobody planned to govern. The gap between deployment speed and oversight capacity keeps widening on both fronts.

And for developers building on top of all this infrastructure, the tooling is moving fast. We dug into Anthropic’s Compaction API, which lets conversations run past the 1M token window—58.6% token reduction in benchmarks, $0.99 per compaction step on Opus 4.6, with a 150K default trigger threshold and a 50K minimum. We also walked through OpenAI’s Agents SDK, now past v0.12 with 20,200+ GitHub stars, where mixing GPT-5.4 with nano models cuts costs from $900/month to $140/month at 1,000 requests per day—a 6x reduction. The frontier labs are competing on developer infrastructure as aggressively as they compete on benchmarks, because the stack you build on today determines the vendor you’re locked into tomorrow.

Latest from PulseMark

How to Build Infinite Conversations with Claude’s Compaction API Threshold tuning, budget enforcement, per-step cost math on Opus 4.6, and the compaction loop failure mode that will find you at 2 AM. Beta since January 12, 2026. Read the tutorial →

OpenAI Agents SDK: Build Multi-Agent Workflows in Python The agents-as-tools pattern from zero to a running orchestrator in under 100 lines, with guardrails, the 5 production pitfalls, and why mixed-model routing is the real cost lever. Read the tutorial →